David's Wikiblog

Mardi, 21. Mai 2013

• PF_RING
• netmap
• netmap prez
• Intel DPDK
• Intel DPDK getting started
• user space network stack
• Direct cache access for high bandwith I/O
• Efficient network applications on multi-core linux

Mardi, 14. Mai 2013

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

Ski (vignettes)

Contenu de la gallerie

Ski (grandes)

Contenu de la gallerie

Tchoupi passe un week end a la maison avant de retourner a l'ecole.

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

Lundi, 13. Mai 2013

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

Lundi, 1. Octobre 2012

List of papers found about floating point hacks

• Christian Plesner Hansen on 0x5f3759df
• Mc Eniry on 0x5f3759df
• Lomont on 0x5f3759df
• Jim blinn's floating point hacks
• Matthew Robertson on 0x5f3759df
• Jan Kadlec on 0x5f3759df
• Chris Miller on 0x5f3759df
• cbloom : fast log and fast exp
• Nicol Schraudolph : fast exp
• Martin ankerl : fast pow
• Jose Fonseca's fast pow
• Harrison Ainsworth's fast pow
• Bruce Dawson tricks part 1
• Bruce Dawson tricks part 2
• Bruce Dawson tricks part 3

One weird thing is that some of those tricks work for x>1, but not for x in [0, 1[.

Tested between [0, 1[

invsqrt

With the following code

 float myinvsqrt(float x)
 {
   union {float f; uint32_t u;} y;
 //   float xhalf = 0.5f * x;
   y.f = x;
   y.u = 0x5f3759df - (y.u >> 1);
 //   y.f = y.f*(1.5f-(xhalf*y.f*y.f));
   return y.f;
 }

I even removed the newton correction ! Here's the graph : inv sqrt approx

Pretty accurate, that one works !

log2

With the following code

 float mylog2(float x)
 {
   union {float f; uint32_t u;} y;
   y.f = x;
   return y.u / 8388608.0 - 127;
 }

Here's the graph : log2 approx

Pretty accurate, that one works !

exp2

With the following code

 float myexp2(float x)
 {
   union {float f; uint32_t u;} y;
   y.u = (x + 127) * 8388608;
   return y.f;
 }

Here's the graph : exp2 approx

Not very accurate, don't know if that works better for x>1.

pow(a, b)

By combining log2 and exp2, we can create an estimate of pow like this :

 float pow(float a, float b)
 {
   union {float f; uint32_t u;} y;
   y.f = a;
   y.u = y.u * b - 127 * (b - 1) * 8388608;
   return y.f;
 }

Example with b = 8.125

pow(x, 8.125) approx

mmmm... not very good. Needs a Newton-Raphson iteration. By adding a single iteration

 y.f = y.f * ((1-8.125) + 8.125 * x * powf(y.f, -1.f/8.125f));

This gives this graph : pow(x, 8.125) approx with newton

So better, only problem, I need powf for the newton iteration... So this is work in progress, hope to find a solution soon.

pow(x, -1/8)

I was able to derive the -1/8 power, here it is

 float mypowm0_125(float x)
 {
   union {float f; uint32_t u;} y;
   y.f = x;
   y.u = 0x476983e4 - (y.u >> 3);
   return y.f;
 }

Here's the graph : -1/8 approx

That one needs a newton iteration to correct for the error.

Mercredi, 30. Mai 2012

Photos en retard d'avril 2012 a San Francisco avec papi et mamie des Charentes

vignettes

Contenu de la gallerie

grandes photos

Contenu de la gallerie

Photos en retard d'Avril 2012 a NY avec papi et mamie des Charentes

vignettes

Contenu de la gallerie

grandes photos

Contenu de la gallerie

Ca y est, la petite puce n'a plus besoin des petites roues.
On dirait presque qu'elle a fait ca toute sa vie.

Papa, il va falloir investir cet ete !

La video.

Mardi, 22. Mai 2012

Ce dimanche, Zoe faisait son spectacle de danse !

Quelques photos

vignettes

Contenu de la gallerie

grandes photos

Contenu de la gallerie

Mardi, 17. Avril 2012

Dr.Dobb's articles by Herb Sutter :

• Free lunch is over by Herb Sutter
• Design for Manycore Systems by Herb Sutter
• Lock-Free Queues by Petru Marginean
• Lock-Free Code: A False Sense of Security by Herb Sutter
• Writing Lock-Free Code: A Corrected Queue by Herb Sutter
• Writing a Generalized Concurrent Queue by Herb Sutter
• Measuring Parallel Performance: Optimizing a Concurrent Queue by Herb Sutter
• Understanding Parallel Performance by Herb Sutter

Dr.Dobb's articles by Andrei Alexandrescu :

• Lock-Free Data Structures by Andrei Alexandrescu
• Lock-Free Data Structures with Hazard Pointers by Andrei Alexandrescu

Articles by Bartosz Milewski :

• Multicores and Publication Safety
• Who ordered memory fences on an x86?
• Who ordered sequential consistency?
• C++ atomics and memory ordering
• The inscrutable c++ memory model

Theoretical papers :

• CAS-Based Lock-Free Algorithm for Shared Deques by Maged M. Michael
• Obstruction-Free Synchronization: Double-Ended Queues as an Example by Maurice Herlihy, Victor Luchangco and Mark Moir
• Lock-free Dynamically Resizable Arrays
• Lock-Free and Practical Deques and Doubly Linked Lists using Single-Word Compare-And-Swap1
• An Optimistic Approach to Lock-Free FIFO Queues
• Fast and Lock-Free Concurrent Priority Queues for Multi-Thread Systems
• Split-Ordered Lists: Lock-Free Extensible Hash Tables
• High Performance Dynamic Lock-Free Hash Tables and List-Based Sets
• Practical lock-freedom
• Real-Time Computing with Lock-Free Shared Objects
• Scalable lock-free dynamic memory allocation
• simple, fast, and practical non blocking and blocking concurrent queue algorithms
• Hazard pointers
• Safe memory reclamation for dynamic lock-free objects using atomic reads and writes
• The repeat offender problem (ROP)
• Using Elimination to Implement Scalable and Lock-Free FIFO Queues
• J. D. Valois. Implementing Lock-Free Queues.
• Correction of a memory management method for lock-free data structures
• Allocating memory in a lock-free manner
• A Scalable Lock-free Stack Algorithm
• A Practical Multi-Word Compare-and-Swap Operation
• Scalable Queue-Based Spin Locks with Timeout
• Mostly Lock-Free Malloc
• Wait free SPSC bounded and unbounded queue with no barrier

others :

• Gcc atomic wiki page
• Lock free containers
• Atomic counter whitepaper
• Futexes
• Mutex and memory visibility
• Memory Barriers: a Hardware View for Software Hackers
• Introduction to lock-free/wait-free and the ABA problem here and here
• non blocking multi producers single consumer queue
• multi producer/multi consumer
• What every programmer should know about memory
• Understanding the linux kernel
• Articles at locklessinc
• Articles at 1024 cores
• Blog at 1024 cores
• Memory Ordering in Modern Microprocessors, Part I
• Memory Ordering in Modern Microprocessors, Part II
• What is RCU
• urcu
• My first RCU design
• Modern micro processors
• cbloom rants
• Flat combining

libraries :

• libsync
• atomic_ops
• gcc atomic builtins
• thread building blocks
• Atomic ptr plus project
• lib lock free data structure
• a scalable allocator : OSlash
• cpp framework
• lock free
• boost began impl of c++0x
• google perf tools
• boost impl of c++0x
• review of available atomic libs
• concurency kit

Mercredi, 21. Mars 2012

Zoe Lise Ewann part 2

Vignettes

Contenu de la gallerie

Grandes photos

Contenu de la gallerie

Zoe-Lise-Ewann part 1/2

Vignettes

Contenu de la gallerie

Grandes photos

Contenu de la gallerie

Photos en retard noel

vignettes

Contenu de la gallerie

grand format

Contenu de la gallerie

Photos en retard noel

vignettes

Contenu de la gallerie

grand format

Contenu de la gallerie

Photos en retard de l'anniv des petits

vignettes

Contenu de la gallerie

photos

Contenu de la gallerie

• Se logguer

• 
• 

• Temps d'éxécution PHP : 0.239 secondes

• Temps d'éxécution SQL : 0.015 secondes